Anthropic has introduced its latest AI model, Claude Mythos Preview, surpassing its previous Claude Opus 4.6 model and competitors like Gemini 3.1 on various benchmarks. However, due to cybersecurity concerns, Anthropic has delayed the release of Claude Mythos Preview, which has demonstrated exceptional power by identifying numerous cybersecurity vulnerabilities missed by humans across applications and operating systems.
The team led by Dario Amodei revealed that Claude Mythos Preview uncovered thousands of zero-day vulnerabilities in web browsers and operating systems, posing significant cybersecurity risks. Notably, it detected a 27-year-old vulnerability in OpenBSD and multiple vulnerabilities in the Linux kernel. These vulnerabilities could potentially allow an attacker to gain complete control of a system.
Originally discovered in a data leak of Anthropic’s CMS, Claude Mythos was anticipated to be the company’s most potent AI model. In response to these findings, Anthropic has initiated Project Glasswing to collaborate with leading tech companies in identifying and patching such vulnerabilities across various platforms.
The consortium includes major players like Apple, Microsoft, Google, Amazon Web Services, CrowdStrike, Palo Alto Networks, Cisco, Broadcom, and the Linux Foundation. Anthropic has allocated up to $100 million in model usage credits to support Project Glasswing, with partners covering additional expenses.
Anthropic emphasizes the importance of addressing vulnerabilities now, as advanced AI models like Claude Mythos become more prevalent. The company highlights that Claude Mythos, not specifically designed for cybersecurity, benefits from its advanced coding and reasoning capabilities as a general-purpose model.
Furthermore, Anthropic is working with U.S. government agencies and security organizations to establish guidelines for deploying such powerful models responsibly. Earlier this year, the startup clashed with the US Pentagon over AI usage disagreements, leading to the termination of a defense deal.
In conclusion, Anthropic’s efforts underscore the necessity of proactive cybersecurity measures in the face of advancing AI capabilities, aiming to create a more secure digital landscape.